A new report from security firm Eclypsium highlights that threat actors can sabotage AI models by tampering with the underlying hardware and firmware of the servers running them. Unlike direct data poisoning, subtle runtime changes can manipulate an AI model to weaken certificate validation or intentionally generate insecure code dependencies. Standard model integrity checks often fail to catch these exploits because the compromise occurs deep within the supporting software and memory controllers rather than the model weights themselves. The findings underscore an underappreciated attack vector targeting high-performance computing components, such as GPUs, accelerators, and enterprise server gear. Cybersecurity experts urge frontier labs and corporate adopters to expand their risk frameworks to include continuous firmware verification and hardware supply chain monitoring.

Full report : Security researchers reveal how low-level firmware and hardware manipulation can silently trigger massive data poisoning.